Distinguishing between authentic and fictitious user-generated hotel reviews

摘要

The objective of this paper is to distinguish between authentic and fictitious user-generated hotel reviews. To achieve this objective, it adopts a two-step approach. The first seeks to classify authentic and fictitious reviews by leveraging on their possible textual differences. The second step attempts to identify the textual traits that are unique to authentic and fictitious reviews. For the purpose of this paper, a ground truth dataset of 1,800 reviews, uniformly divided between authentic and fictitious, was created. With respect to the first step, authentic and fictitious reviews were classified by using four forms of textual differences: understandability, level of details, writing style, and cognition indicators. Classification was performed using voting by average probability among logistic regression, C4.5, Support Vector Machine, JRip, and Random Forest classifiers. Using five-fold cross-validation, the proposed approach was found to outperform two existing baselines. Furthermore, with respect to the second step, the textual traits unique to authentic and fictitious reviews were identified using Information Gain, and Chi-squared feature selection techniques. A sequential forward feature selection approach was further adopted to identify the top five features that aid the classification of authentic and fictitious reviews. These include the use of nouns, articles, function words, punctuations, and in particular, exclamation points in reviews. The implications of the results are discussed.